Open-Weight Models Drive AI Innovation on Amazon Bedrock
The increasing availability of open-weight models is transforming the economics of developing and deploying artificial intelligence solutions at scale. This trend allows companies to select the optimal balance of capability, speed, and cost for various workloads. Amazon Web Services (AWS) is positioning itself to support an era where organizations can adopt open-source innovation while maintaining the high level of security and reliability required for production environments.
Today, Moonshot AI has released Kimi K3, making it available on Amazon Bedrock. This addition provides users with a powerful new option for complex coding and knowledge-based tasks. According to Moonshot AI, Kimi K3 represents the company’s most capable model and is notable as the first open model to achieve 2.8 trillion parameters. The model integrates native vision capabilities and features a 1-million-token context window, offering an estimated 2.5x improvement in scaling efficiency compared to its predecessor, Kimi K2. These advancements make Kimi K3 highly suitable for prolonged coding and knowledge workflows that require sustained context across large datasets, documents, and images.
AWS Investment and Security Features
The introduction of Kimi K3 highlights AWS’s continuous commitment to open-weight models within the Amazon Bedrock ecosystem. Since 2025, Bedrock has expanded its catalog by incorporating numerous open-weight models from providers such as DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, and Qwen.
AWS has also continually advanced the underlying inference technology supporting these models. In 2026, the platform added crucial capabilities, including tool calling, structured output generation, reasoning, response streaming, and dedicated Responses and Chat Completions APIs. Because these features are platform-level enhancements rather than model-specific integrations, newly available open-weight models can benefit from them immediately.
Maintaining Data Control with Open Models
Users can adopt Kimi K3 on Amazon Bedrock without altering their existing security protocols. All data processing occurs within the AWS data boundary. The platform guarantees that user data is neither shared with the model provider nor used to train the underlying model. Furthermore, zero data retention is always enabled for all inference requests, and zero operator access prevents AWS personnel from viewing prompts or completions during inference, allowing users to utilize open-weight models with confidence while retaining full control over their data.
Accessing and Optimizing Kimi K3
Users can begin testing Kimi K3 directly within the Amazon Bedrock console by navigating to Test > Playground and selecting the model. For programmatic use, the model can be called via the `bedrock-runtime` endpoint, which supports the OpenAI-compatible Responses and Chat Completions APIs, alongside the Amazon Bedrock Invoke and Converse API APIs.
When running workloads, users have two main options for cross-Region inference profiles. For tasks without geographical limitations, the global profile, `global.moonshotai.kimi-k3`, is recommended, as it routes requests to any supported commercial AWS Region worldwide and costs approximately 10% less than a geographic profile. Conversely, the US geographic profile, `us.moonshotai.kimi-k3`, should be used when data residency requirements mandate processing within the United States.
Optimizing Inference with Explicit Prompt Caching
For long-running coding and knowledge workflows, stable context—such as system instructions, tool definitions, or reference documents—is often resent. Explicit prompt caching allows developers to identify and save reusable prompt prefixes. When a subsequent request utilizes a cached prefix, Amazon Bedrock can reduce input token costs and decrease response latency.
To implement caching for Kimi K3 on Amazon Bedrock, users must add a `prompt_cache_breakpoint` to the end of a reusable prompt prefix (which must exceed 1,024 tokens). In explicit mode, the initial tokens written to the cache are billed at a higher rate but are retained for at least 30 minutes. For matching subsequent requests that hit the cache, input tokens are billed at a reduced rate and will not consume input-tokens-per-minute quotas.
When using the OpenAI Python API, explicit caching can be configured as follows:
from aws_bedrock_token_generator import provide_token
from openai import OpenAIregion = “us-west-2”
oai_client = OpenAI(
api_key=provide_token(region=region),
base_url=f”https://bedrock-runtime.{region}.amazonaws.com/openai/v1″,
)resp = oai_client.responses.create(
input=”What is Byte-Pair Encoding, in AI?”,
model=”global.moonshotai.kimi-k3″,
)
print(resp.output_text)
The detailed configuration for explicit caching using the OpenAI Python API is shown below:
resp = oai_client.responses.create(
model=”global.moonshotai.kimi-k3″,
# Enable explicit caching mode:
extra_body={“prompt_cache_options”: {“mode”: “explicit”}},
input=[
{
“type”: “message”,
“role”: “system”,
“content”: [
{
“type”: “input_text”,
“text”: SYSTEM_PROMPT,
# A long, static system prompt is a great target for caching:
“prompt_cache_breakpoint”: {“mode”: “explicit”},
},
]
},
{
“type”: “message”,
“role”: “user”,
“content”: [
{
“type”: “input_text”,
“text”: USER_INPUT,
# Multiple breakpoints can also be defined, for layered cache:
“prompt_cache_breakpoint”: {“mode”: “explicit”},
},
],
},
],
)
if resp.usage.input_tokens_details.cached_tokens:
print(“Hit cache!”)Implementing Kimi K3 in Practical Workflows
Beyond direct API calls, Kimi K3 can be integrated into various coding assistants, personal agents, and agentic frameworks that support Amazon Bedrock or general OpenAI-compatible model providers.
Coding Assistants
Tools like OpenCode, an open-source, model-agnostic platform, provide a native Amazon Bedrock model provider that utilizes the Converse API. Developers can configure the `amazon-bedrock` provider either at the user or project level. Once configured, OpenCode automatically detects available Amazon Bedrock models, allowing users to select `global.moonshotai.kimi-k3` using the `/models` command.
An example configuration file for OpenCode using the Amazon Bedrock provider is structured as follows:
{
“$schema”: “https://opencode.ai/config.json”,
“model”: “amazon-bedrock/global.moonshotai.kimi-k3”,
“provider”: {
“amazon-bedrock”: {
“options”: {
“region”: “us-west-2”,
“profile”: “PLACEHOLDER-YOUR-AWS-PROFILE-NAME”
}
}
}
}Productivity Agents
For general productivity, open-source assistants such as Hermes Agent can be utilized. Hermes supports use cases like deep research and task automation, areas where Kimi K3 performs effectively. Hermes natively supports models on Amazon Bedrock. To connect the agent to Kimi K3, users must:
- Run `hermes model` in the terminal.
- Select “AWS Bedrock” from the list of providers.
- Choose a credential method (default credential chain is recommended) or generate an Amazon Bedrock API key.
- Select Kimi K3 from the auto-discovered model list, or manually enter `global.moonshotai.kimi-k3`.
With the provider and model configured, Kimi K3 can be used for complex agentic workflows, such as building a personalized study plan.
Availability and Next Steps
Kimi K3 is available immediately on Amazon Bedrock through both the US Geo (us.) and Global (global.) cross-Region inference profiles. For comprehensive details on supported Regions and pricing, users should consult the official Amazon Bedrock documentation.
Users are encouraged to test the model in the Amazon Bedrock console or explore the resources available in the Moonshot AI on AWS samples repository on GitHub.
